Enhancing anti-inflammatory small molecules prediction using transformer embeddings and hybrid features integration with ensemble models

Inflammation is a complex biological response that helps the body defend itself against toxins, pathogens, and injuries. However, prolonged inflammation can lead to chronic disorders such as cardiovascular diseases and neurodegenerative conditions etc. Developing effective anti-inflammatory small molecule drugs is essential for the management of all these conditions, but traditional methods are time-consuming and resource intensive. To accelerate this process, machine learning and deep learning approaches are widely used to predict the potential anti-inflammatory molecules based on molecular features. In this study, we used eight types of molecular representations, including fingerprints and descriptors such as Morgan, MACCS, and RDKit, as well as hybrid features like Morgan_RDKit, Morgan_MACCS, and MACCS_RDKit. Additionally, we incorporated transformer-based embeddings generated by Llama 2 and ChemBERTa. In total, 24 models were developed with Stacking Classifier on the hybrid features (MACCS_RDKit), achieving the highest performance, with an accuracy of 0.95, Mathews Correlation Coefficient (MCC) of 0.89, and sensitivity of 0.95 on an independent test dataset. These results highlight the effectiveness of hybrid feature representations combined with ensemble learning in enhancing predictive performance for anti-inflammatory small molecules. Moreover, SHAP analysis revealed consistent key descriptors and structural fragments that significantly impact the model's predictions.